language:
- en
pretty_name: GoHuntJBDS
size_categories: - 1K<n<10K
Dataset Card for GoHuntJBDS
Dataset Summary
GoHuntJBDS (Go Hunt JailBreak DataSet) is a curated dataset of adversarial prompts and their model responses, intended for the evaluation of jailbreak behavior in large language models (LLMs). It includes both safe and unsafe outputs generated by various LLMs, with labels indicating whether each response constitutes a jailbreak. The dataset supports research on model robustness, alignment, safety auditing, and adversarial defense mechanisms.
Supported Tasks and Leaderboards
- Jailbreak Detection: Classify whether a model response represents a successful jailbreak.
- Adversarial Prompt Evaluation: Evaluate the robustness of LLMs against adversarial input.
- Safety Filter Benchmarking: Measure how well models or classifiers can detect and refuse unsafe content.
Languages
- English
Dataset Structure
Data Fields
Each row in GoHuntJBDS.csv contains:
prompts(string): The adversarial or baseline prompt given to the model.response(string): The corresponding model-generated completion.response_type(int): Binary label — 1 for unsafe (jailbroken), 0 for safe.prompt_category(string): The content domain of the prompt (e.g., Economic Harm, Malware).prompt_source(string): The origin of the prompt (e.g., AdvBench, GPT Gen_